XB+-Tree: A Novel Index for PCM/DRAM-Based Hybrid Memory

Abstract

Phase Change Memory (PCM) has emerged as a new kind of future memories that can be used as an alternative of DRAM. PCM has a number of special properties such as non-volatility, high density, read/write asymmetry, and byte addressability. Specially, PCM has higher write latency than DRAM but has comparable read latency with DRAM. This makes it difficult to directly replace DRAM with PCM in current memory hierarchy. Thus, in this paper, we propose to construct hybrid memory architecture that involves both PCM and DRAM, which is a practical and feasible way to utilize PCM. Such hybrid memory architecture introduces many new issues for database researches, as existing algorithms have to be revised to be suitable for hybrid memory. In this paper, we study the indexing issue on PCM/DRAM-based hybrid memory and propose an improved version of the B+-tree called XB+-tree (eXtended B+-tree). The key idea of the XB+-tree is to detect the read/write tendency of the nodes in the tree index and organize write-intensive nodes on PCM while putting read-intensive nodes on DRAM. We propose a new node management and migration algorithm in the XB+-tree to effectively move nodes between DRAM and PCM. With this mechanism, we can reduce the read and write operations on PCM and improve the overall performance. We conduct trace-driven experiments and compare our proposal with three existing indices including the B+-tree, the OB+-tree (B+-tree with the overflow scheme), and the CB+-tree. The results in terms of PCM read/write count and run time suggest the efficiency of our proposal.